One of my test prompts asked Claude Code to "Write pytest tests for this Python function, with a fixture for the sample data", followed by a one-line median function. The workspace had the public python-testing-patterns skill, whose description says to use it when writing Python tests, names pytest and fixtures, and shares five words with the prompt. I ran the prompt 18 times, three on each of Haiku, Sonnet and Opus with the skill's original description and three with a "pushy" rewrite of it. The skill never loaded. Sonnet started writing a test file in all six of its runs, Opus saved the function to its own file in five, and Haiku listed the empty folder and checked for pytest in five. Claude Code reports nothing on a miss like that; you only see it by checking the transcript for a Skill call.

Anthropic's skill-creator skill says Claude tends to "undertrigger" skills and suggests making descriptions a little "pushy" [1].…